An algorithm for traffic flow parameters estimation and prediction using composition of machine learning methods and time series models

A problem of traffic flow analysis and prediction in city transport network is considered in this paper. The proposed algorithm uses GPS / GLONASS data of public transport location as input. Projecting this information on a transport network graph, as well as using additional filtering, we estimate traffic flow parameters. These parameters are used for short-term (up to 1 hour) prediction of road conditions in the city transport network. There is proposed a new method which consists of several steps to construct prediction. First, the transport graph is divided into a number of subgraphs by a territorial basis. Second, we use a dimension reduction method based on principal components analysis to describe the spatio-temporal distribution of traffic flow condition in the subgraphs. Third, an elementary prediction for each of the subgraphs is formed using the potential functions method with the measure of the subgraphs descriptions closeness introduced by analogy with bilateral filtering and support vector machine. Fourth, the additional elementary prediction is calculated using the known scalar and vector Box-Jenkins time series prediction models. Fifth, we construct the result prediction for each of the subgraphs using an adaptive linear composition of elementary predictions. At last, the traffic flow parameters are calculated as a linear combination of predictions for subgraphs of the city transport network. We have also made experimental investigations of transport network in Samara to evaluate the prediction accuracy of the proposed algorithm. The advantages of the proposed solution in comparison with existing ones are provided.